The existing code for reading and writing the return value can
overflow the passed in buffers in a couple of situations. This commit
aims to resolve these issues.
The problems were detected using valgrind, here are two examples,
first from gdb.base/structs.exp:
(gdb) p/x fun9()
==31353== Invalid write of size 8
==31353== at 0x4C34153: memmove (vg_replace_strmem.c:1270)
==31353== by 0x632EBB: memcpy (string_fortified.h:34)
==31353== by 0x632EBB: readable_regcache::raw_read(int, unsigned char*) (regcache.c:538)
==31353== by 0x659D3F: riscv_return_value(gdbarch*, value*, type*, regcache*, unsigned char*, unsigned char const*) (riscv-tdep.c:2593)
==31353== by 0x583641: get_call_return_value (infcall.c:448)
==31353== by 0x583641: call_thread_fsm_should_stop(thread_fsm*, thread_info*) (infcall.c:546)
==31353== by 0x59BBEC: fetch_inferior_event(void*) (infrun.c:3883)
==31353== by 0x53890B: check_async_event_handlers (event-loop.c:1064)
==31353== by 0x53890B: gdb_do_one_event() [clone .part.4] (event-loop.c:326)
==31353== by 0x6CA34B: wait_sync_command_done() (top.c:503)
==31353== by 0x584653: run_inferior_call (infcall.c:621)
...
And from gdb.base/call-sc.exp:
(gdb) advance fun
fun () at /gdb/gdb/testsuite/gdb.base/call-sc.c:41
41 return foo;
(gdb) finish
==1968== Invalid write of size 8
==1968== at 0x4C34153: memmove (vg_replace_strmem.c:1270)
==1968== by 0x632EBB: memcpy (string_fortified.h:34)
==1968== by 0x632EBB: readable_regcache::raw_read(int, unsigned char*) (regcache.c:538)
==1968== by 0x659D01: riscv_return_value(gdbarch*, value*, type*, regcache*, unsigned char*, unsigned char const*) (riscv-tdep.c:2576)
==1968== by 0x5891E4: get_return_value(value*, type*) (infcmd.c:1640)
==1968== by 0x5892C4: finish_command_fsm_should_stop(thread_fsm*, thread_info*) (infcmd.c:1808)
==1968== by 0x59BBEC: fetch_inferior_event(void*) (infrun.c:3883)
==1968== by 0x53890B: check_async_event_handlers (event-loop.c:1064)
==1968== by 0x53890B: gdb_do_one_event() [clone .part.4] (event-loop.c:326)
==1968== by 0x6CA34B: wait_sync_command_done() (top.c:503)
...
There are a couple of problems with the existing code, that are all
related.
In riscv_call_arg_struct we incorrectly rounded up the size of a
structure argument. This is unnecessary, and caused GDB to read too
much data into the output buffer when extracting a struct return
value.
In fixing this it became clear that we were incorrectly assuming that
any value being placed in a register (or read from a register) would
always access the entire register. This is not true, for example a
9-byte struct on a 64-bit target places 8-bytes in one registers and
1-byte in a second register (assuming available registers). To handle
this I switch from using cooked_read to cooked_read_part.
Finally, when processing basic integer return value types these are
extended to xlen sized types and then passed in registers. We
currently don't handle this type expansion in riscv_return_value, but
we do in riscv_push_dummy_call. The result is that small integer
types (like char) result in a full xlen sized register being written
into the output buffer, which results in buffer overflow. To address
this issue we now create a value of the expanded type and use this
values contents buffer to hold the return value before casting the
value down to the smaller expected type.
This patch resolves all of the valgrind issues I have found so far,
and causes no regressions. Tested against RV32/64 with and without
floating point support.
gdb/ChangeLog:
* riscv-tdep.c (riscv_call_arg_struct): Don't adjust size before
assigning locations.
(riscv_return_value): Take more care not to read/write outside of
argument buffer. Cast return value between the declared type and
the abi type.
We should save and restore the floating point status registers. This
became an issue when testing 32-bit float on a target with 64-bit with
the gdb.base/callfuncs.exp test.
gdb/ChangeLog:
* riscv-tdep.c (riscv_register_reggroup_p): Save and restore fcsr,
fflags, and frm registers.
Provide a mapping between GDB's register numbers and DWARF's register
numbers. This resolves some failures that I was seeing on
gdb.base/store.exp when running on an rv64imfdc target.
gdb/ChangeLog:
* riscv-tdep.c (riscv_dwarf_reg_to_regnum): New function.
(riscv_gdbarch_init): Register new function with gdbarch.
* riscv-tdep.h: New enum to define RISC-V DWARF register numbers.
While discussing this issue:
https://sourceware.org/ml/gdb-patches/2018-12/threads.html#00082
I added a printf gated by "set debug symtab-create" to be able to
quickly see all minimal symbols recorded by GDB. I thought it would be
useful to have it built-in, for the future. Here's how the output
looks:
Recording minsym: mst_data 0x400780 15 _IO_stdin_used
Recording minsym: mst_text 0x400700 13 __libc_csu_init
Recording minsym: mst_bss 0x601058 25 _end
gdb/ChangeLog:
* minsyms.c (mst_str): New.
(minimal_symbol_reader::record_full): Add debug output.
Various tests use test code written in i386 / x86_64 assembly that cannot
be used to create PIE executables. Therefore compilation of test programs
failed on systems where the compiler default is to create PIE executable.
The solution is to use -no-pie linker flag, however, such flag may not
(is not) supported by all compilers GDB needs to support (e.g. gcc 4.8).
To handle this, introduce a new flag to gdb_compile - nopie - which
inserts -no-pie linker flag where supported and is no-op where it is
not. By default, -no-pie flag is inserted since most modern compiler do
support it.
The ptrace command PT_LWPINFO to request detailed information about a
stopped thread can return stale signal information from an earlier
stop. Events which are reporting an intercepted signal will always
report the correct information, but signal stops for some other events
such as system call enter/exit events might include stale siginfo from
an earlier signal. In particular, if a thread reports a system call
entry or exit event after previously reporting a single-step or
breakpoint event via SIGTRAP, fbsd_handle_debug_trap believed the
system call event was the previous event and claimed it resulting in a
spurious SIGTRAP event.
True breakpoint and single-step events will never report another event
in the pl_flags member of struct ptrace_lwpinfo. Use this to detect
stale siginfo by requiring pl_flags to have only the PL_FLAG_SI flag
and no other flags before treating a SIGTRAP as a single-step or
breakpoint trap.
gdb/ChangeLog:
* fbsd-nat.c (fbsd_handle_debug_trap): Require pl.pl_flags to
equal PL_FLAG_SI.
(fbsd_nat_target::stopped_by_sw_breakpoint): Likewise.
Running `info os someUnknownOsType` is crashing when gdb is built with
-D_GLIBCXX_DEBUG:
/usr/include/c++/5/debug/vector:439:error: attempt to
access an element in an empty container.
In target_read_stralloc from target.c, the call to
target_read_alloc_1 can return an empty vector, we then call vector::back on
this vector, which is invalid.
This commit adds a check for emptiness before trying to call
vector::back on it. It also adds test to check for `info os <unknown>`
to return the proper error message.
This is a regression in gdb 8.2 and this patch restores the behavior of
previous versions.
gdb/ChangeLog:
PR gdb/23974
* target.c (target_read_stralloc): Check for empty vector.
gdb/testsuite/ChangeLog:
PR gdb/23974
* gdb.base/info-os.exp: Check return for unknown "info os" type.
When using the --configuration command line switch, or using
the "show configuration" command with a version of GDB which
was configured without Python supoprt, this patch changes
the resulting output to include...
--without-python
... instead of not printing anything about Python support.
gdb/ChangeLog:
* top.c (print_gdb_configuration): Print "--without-python"
if GDB was configured without Python.
Tested on x86_64-linux by rebuilding GDB with and without Python,
and checking the output of "gdb --configuration" in both cases.
Avoid compiler errors caused by trying to print CORE_ADDR using '%ld'
format, instead convert to a string and print that instead.
gdb/ChangeLog:
* riscv-tdep.c (riscv_scan_prologue): Use plongest to format
a signed offset as a string.
When compiling with clang 3.8 (default clang version on Debian
Stretch, the current stable), we get errors like this:
CXX dtrace-probe.o
../../binutils-gdb/gdb/dtrace-probe.c:103:31: error: default initialization of an object of const type 'const dtrace_static_probe_ops' without a user-provided default constructor
const dtrace_static_probe_ops dtrace_static_probe_ops;
^
Silence them by value-initializing those objects. It's not necessary
with other compilers (later clang versions, gcc), but it shouldn't
hurt either.
* elf32-i386.c (elf_i386_relocate_section): Call
elf_i386_rtype_to_howto to get reloc_howto_type pointer.
* elf64-x86-64.c (elf_x86_64_relocate_section): Call
elf_x86_64_rtype_to_howto to get reloc_howto_type pointer.
standard_ada_testfile, standard_test_file and the explicit
csrcfile assignment in info_auto_lang.exp all gives similar pathnames
prefix for a source, such as
/home/philippe/gdb/git/build_binutils-gdb/gdb/testsuite/../../../binutils-gdb/gdb/testsuite/gdb.<something>.
Note that the above pathnames contain ../ which appears when a relative
pathname is used to call configure.
In any case, the gnat compiler normalizes Ada sources path when compiling.
So, the 'Ada' .o object are referencing a pathname such as
/home/philippe/gdb/git/binutils-gdb/gdb/testsuite/gdb.ada/info_auto_lang/proc_in_ada.adb,
while the 'C' .o object still references the not normalized pathname.
As the results of 'info functions | ...' are sorted by pathname first,
the order of the results depends on the comparison between different directories,
leading to results that can change depending on these directories.
=> Ensure the result order is always the same, by normalising the C source file,
which makes the results independent of the way configure is launched.
Tested by running the testcase in 2 different builds, that without normalize
were giving different results.
Note: such 'set csrcfile' is used in 4 other tests mixing Ada and C.
After discussion, it was deemed sufficient to just normalize the pathname
for this test.
gdb/testsuite/ChangeLog
2018-12-20 Philippe Waroquiers <philippe.waroquiers@skynet.be>
* gdb.ada/info_auto_lang.exp: Normalize some_c source file.
Update order of results accordingly.
Supply default gdbarch methods for gdbarch_dummy_id,
gdbarch_unwind_pc, and gdbarch_unwind_sp. This patch doesn't actually
convert any targets to use these methods, and so, there will be no
user visible changes after this commit.
The implementations for default_dummy_id and default_unwind_sp are
fairly straight forward, these just take on the pattern used by most
targets. Once these default methods are in place then most targets
will be able to switch over.
The implementation for default_unwind_pc is also fairly straight
forward, but maybe needs some explanation.
This patch has gone through a number of iterations:
https://sourceware.org/ml/gdb-patches/2018-03/msg00165.htmlhttps://sourceware.org/ml/gdb-patches/2018-03/msg00306.htmlhttps://sourceware.org/ml/gdb-patches/2018-06/msg00090.htmlhttps://sourceware.org/ml/gdb-patches/2018-09/msg00127.html
and the implementation of default_unwind_pc has changed over this
time. Originally, I took an implementation like this:
CORE_ADDR
default_unwind_pc (struct gdbarch *gdbarch, struct frame_info *next_frame)
{
int pc_regnum = gdbarch_pc_regnum (gdbarch);
return frame_unwind_register_unsigned (next_frame, pc_regnum);
}
This is basically a clone of default_unwind_sp, but using $pc. It was
pointed out that we could potentially do better, and in version 2 the
implementation became:
CORE_ADDR
default_unwind_pc (struct gdbarch *gdbarch, struct frame_info *next_frame)
{
struct type *type;
int pc_regnum;
CORE_ADDR addr;
struct value *value;
pc_regnum = gdbarch_pc_regnum (gdbarch);
value = frame_unwind_register_value (next_frame, pc_regnum);
type = builtin_type (gdbarch)->builtin_func_ptr;
addr = extract_typed_address (value_contents_all (value), type);
addr = gdbarch_addr_bits_remove (gdbarch, addr);
release_value (value);
value_free (value);
return addr;
}
The idea was to try split out some of the steps of unwinding the $pc,
steps that are on some (or many) targets no-ops, and so allow targets
that do override these methods, to make use of default_unwind_pc.
This implementation remained in place for version 2, 3, and 4.
However, I realised that I'd made a mistake, most targets simply use
frame_unwind_register_unsigned to unwind the $pc, and this throws an
error if the register value is optimized out or unavailable. My new
proposed implementation doesn't do this, I was going to end up
breaking many targets.
I considered duplicating the code from frame_unwind_register_unsigned
that throws the errors into my new default_unwind_pc, however, this
felt really overly complex. So, what I instead went with was to
simply revert back to using frame_unwind_register_unsigned. Almost
all existing targets already use this. Some of the ones that don't can
be converted to, which means almost all targets could end up using the
default.
One addition I have made over the version 1 implementation is to add a
call to gdbarch_addr_bits_remove. For most targets this is a no-op,
but for a handful, having this call in place will mean that they can
use the default method. After all this, the new default_unwind_pc now
looks like this:
CORE_ADDR
default_unwind_pc (struct gdbarch *gdbarch, struct frame_info *next_frame)
{
int pc_regnum = gdbarch_pc_regnum (gdbarch);
CORE_ADDR pc = frame_unwind_register_unsigned (next_frame, pc_regnum);
pc = gdbarch_addr_bits_remove (gdbarch, pc);
return pc;
}
gdb/ChangeLog:
* gdb/dummy-frame.c (default_dummy_id): Defined new function.
* gdb/dummy-frame.h (default_dummy_id): Declare new function.
* gdb/frame-unwind.c (default_unwind_pc): Define new function.
(default_unwind_sp): Define new function.
* gdb/frame-unwind.h (default_unwind_pc): Declare new function.
(default_unwind_sp): Declare new function.
* gdb/frame.c (frame_unwind_pc): Assume gdbarch_unwind_pc is
available.
(get_frame_sp): Assume that gdbarch_unwind_sp is available.
* gdb/gdbarch.c: Regenerate.
* gdb/gdbarch.h: Regenerate.
* gdb/gdbarch.sh: Update definition of dummy_id, unwind_pc, and
unwind_sp. Add additional header files to be included in
generated file.
For PLT expressions, we should subtract the PLT relocation size only for
jump instructions. Since PLT relocations are PC relative, we only allow
"symbol@PLT" in PLT expression.
gas/
PR gas/23997
* config/tc-i386.c (x86_cons): Check for invalid PLT expression.
(md_apply_fix): Subtract the PLT relocation size only for jump
instructions.
* testsuite/gas/i386/reloc32.s: Add test for invalid PLT
expression.
* testsuite/gas/i386/reloc64.s: Likewise.
* testsuite/gas/i386/ilp32/reloc64.s: Likewise.
* testsuite/gas/i386/reloc32.l: Updated.
* testsuite/gas/i386/reloc64.l: Likewise.
* testsuite/gas/i386/ilp32/reloc64.l: Likewise.
ld/
PR gas/23997
* testsuite/ld-i386/i386.exp: Run PR gas/23997 test.
* testsuite/ld-x86-64/x86-64.exp: Likewise.
* testsuite/ld-x86-64/pr23997a.s: New file.
* testsuite/ld-x86-64/pr23997b.c: Likewise.
* testsuite/ld-x86-64/pr23997c.c: Likewise.
A recent patch [1] to fix a GCC PR [2] actually broke the GDB build.
To fix, remove the stack pointer clobber. GCC will ignore the clobber
marker, and will not save or restore the stack pointer.
I ran "make check-gdb" on x86_64 to ensure there are no regressions.
gdb/ChangeLog:
2018-12-17 Dimitar Dimitrov <dimitar@dinux.eu>
* nat/linux-ptrace.c (linux_ptrace_test_ret_to_nx): Remove sp clobbers.
[1] https://gcc.gnu.org/ml/gcc-patches/2018-12/msg00532.html
[2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=52813
Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
This test checks code layout by function symbol ordering, but that
doesn't work on powerpc64 ELFv1 where the function symbol is on a
descriptor. A simple work-around is to have nm emit synthetic symbols
marking the code entry point of functions. Since the text segment is
laid out before the data segment, the synthetic symbols will have
lower addresses than function descriptor symbols and be seen first in
nm -n output.
On other targets, nm --synthetic typically emits symbols on plt
entries. Since the testcase doesn't call any of the functions of
interest there shouldn't be plt entries for those functions, so there
should be no potentially confusing extra symbols.
* testsuite/Makefile.am (keep_text_section_prefix_nm.stdout):
Pass --synthetic to nm.
* testsuite/Makefile.in: Regenerate.
All of the backend relocate_section functions that interpret reloc
numbers assuming the input file is of the expected type (ie. same as
output or very similar) really ought to be checking input file type.
Not many do, and those that do currently just assert. This patch
replaces the assertion with a more graceful exit.
PR 23980
* elf32-i386.c (elf_i386_relocate_section): Exit with wrong format
error rather than asserting input file is as expected.
* elf32-s390.c (elf_s390_relocate_section): Likewise.
* elf32-sh.c (sh_elf_relocate_section): Likewise.
* elf32-xtensa.c (elf_xtensa_relocate_section): Likewise.
* elf64-ppc.c (ppc64_elf_relocate_section): Likewise.
* elf64-s390.c (elf_s390_relocate_section): Likewise.
* elf64-x86-64.c (elf_x86_64_relocate_section): Likewise.
* elf32-ppc.c (ppc_elf_relocate_section): Exit with wrong format
error if input file is not ppc32 ELF.
The host syscall callback mechanism should take care of updating the
errcode within the CB_SYSCALL struct, and we should not be adjusting
the error code once the syscall has completed. We especially, should
not be rewriting the syscall errcode based on the value of errno some
time after running the host syscall, as there is no guarantee that
errno has not be overwritten.
To perform a syscall we call cb_syscall (in syscall.c). To return
from cb_syscall control passes through one of two exit paths these are
labeled FinishSyscall and ErrorFinish and are reached using goto
statements scattered throughout the cb_syscall function.
In FinishSyscall we store the syscall result in 'sc->result', and the
error code is transated to target encoding, and stored in
'sc->errcode'.
In ErrorFinish, we again store the syscall result in 'sc->result', and
fill in 'sc->errcode' by fetching the actual errno from the host with
the 'cb->get_errno' callback.
In both cases 'sc->errcode' will have been filled in with an
appropriate value.
Further, if we look at a specific syscall example, CB_SYS_open, in
this case the first thing we do is fetch the path to open from the
target with 'get_path', if this fails then the errcode is returned,
and we jump to FinishSyscall. Notice that in this case, no host
syscall may have been performed, for example a failure to read the
path to open out of simulated memory can return EINVAL without
performing any host syscall. Given that no host syscall has been
performed, reading the host errno makes absolutely no sense.
This commit removes from sim_syscall_multi the rewriting of
sc->errcode based on the value of errno, and instead relies on the
value stored in the cb_syscall.
sim/common/ChangeLog:
* sim-syscall.c (sim_syscall_multi): Don't update sc->errcode at
this point, it should have already been set in cb_syscall.
In the dwarf reader we have a set of predicates, these include the
different producer predicates and also some control predicates. The
older ones are declared as integers, while newer ones (added since the
C++ conversion) are bool.
This commit makes them all bool for consistency. There should be no
user visible change after this commit.
gdb/ChangeLog:
* dwarf2read.c (struct dwarf2_cu): Convert the fields 'mark',
'has_loclist', 'checked_producer', 'producer_is_gxx_lt_4_6',
'producer_is_gcc_lt_4_3', 'producer_is_icc_lt_14',
'processing_has_namespace_info' from unsigned int to bool. Update
comments.
(producer_is_icc_lt_14): Update return type.
(producer_is_gcc_lt_4_3): Likewise.
(producer_is_gxx_lt_4_6): Likewise.
(process_die): Write true instead of 1 into predicate fields.
(dwarf2_start_symtab): Likewise.
(var_decode_location): Likewise.
(dwarf2_mark_helper): Likewise.
(dwarf2_mark): Likewise.
(dwarf2_clear_marks): Write false instead of 0 into predicate
field.
(dwarf2_cu::dwarf2_cu): Initialise predicate fields to false, not
0.
Including asm/sigcontext.h together with libc headers is not valid. In
general linux headers may not work with libc headers, so mixing them
should be avoided, especially when the linux header defines types that
are also exposed in libc headers.
In case of asm/sigcontext.h glibc happens to work because glibc signal.h
directly includes it, but e.g. in musl libc signal.h replicates the
sigcontext.h definitions in an abi compatible way which are in conflict
with the linux definitions when both headers are included.
Since old linux headers or old libc headers may not have the necessary
definitions, gdb has to replicate the definitions it relies on anyway.
Which is fine since all definitions must be ABI stable. For linux apis
that are not available via libc headers, replicating the definitions in
gdb is the most reliable way to use them.
Note: asm/ptrace.h includes asm/sigcontext.h in some versions of linux
headers, which is just as problematic and should be fixed in linux.
gdb/ChangeLog:
* nat/aarch64-sve-linux-ptrace.h: Include signal.h instead of
asm/sigcontext.h.
2018-12-17 Philippe Waroquiers <philippe.waroquiers@skynet.be>
* nat/linux-ptrace.c (kill_child): Fix ARI warning by removing
warning trailing new line.
PR 23994
* aoutx.h: Include limits.h.
(get_reloc_upper_bound): Detect long overflow and return a file
too big error if it occurs.
* elf.c: Include limits.h.
(_bfd_elf_get_symtab_upper_bound): Detect long overflow and return
a file too big error if it occurs.
(_bfd_elf_get_dynamic_symtab_upper_bound): Likewise.
(_bfd_elf_get_dynamic_reloc_upper_bound): Likewise.
Running the gdb testsuite under Valgrind started to fail after 100+ tests,
due to out of memory caused by lingering processes.
The lingering processes are caused by the combination
of a limitation in Valgrind signal handling when using PTRACE_TRACEME
and a (minor) bug in GDB.
The Valgrind limitation is : when a process is ptraced and raises
a signal, Valgrind will replace the raised signal by SIGSTOP as other
signals are masked by Valgrind when executing a system call.
Removing this limitation seems far to be trivial, valgrind signal
handling is very complex.
Due to this valgrind limitation, GDB linux_ptrace_test_ret_to_nx gets
a SIGSTOP signal instead of the expected SIGTRAP or SIGSEGV.
In such a case, linux_ptrace_test_ret_to_nx does an early return, but
does not kill the child (running under valgrind), child stays in a STOP-ped
state.
These lingering processes then eat the available system memory,
till launching a new process starts to fail.
This patch fixes the GDB minor bug by killing the child in case
linux_ptrace_test_ret_to_nx does an early return.
nat/linux-ptrace.c has 3 different logics to kill a child process.
So, this patch factorizes killing a child in the function kill_child.
The 3 different logics are:
* linux_ptrace_test_ret_to_nx is calling both kill (child, SIGKILL)
and ptrace (PTRACE_KILL, child, ...), and then is calling once
waitpid.
* linux_check_ptrace_features is calling ptrace (PTRACE_KILL, child, ...)
+ my_waitpid in a loop, as long as the waitpid status was WIFSTOPPED.
* linux_test_for_tracefork is calling once ptrace (PTRACE_KILL, child, ...)
+ my_waitpid.
The linux ptrace documentation indicates that PTRACE_KILL is deprecated,
and tells to not use it, as it might return success but not kill the tracee.
The documentation indicates to send SIGKILL directly.
I suspect that linux_ptrace_test_ret_to_nx calls both kill and ptrace just
to be sure ...
I suspect that linux_check_ptrace_features calls ptrace in a loop
to bypass the PTRACE_KILL limitation.
And it looks like linux_test_for_tracefork does not handle the PTRACE_KILL
limitation.
Also, 2 of the 3 logics are calling my_waitpid, which seems better,
as this is protecting the waitpid syscall against EINTR.
So, the logic in kill_child is just using kill (child, SIGKILL)
+ my_waitpid, and then does a few verifications to see everything worked
accordingly to the plan.
Tested on Debian/x86_64.
2018-12-16 Philippe Waroquiers <philippe.waroquiers@skynet.be>
* nat/linux-ptrace.c (kill_child): New function.
(linux_ptrace_test_ret_to_nx): Use kill_child instead of local code.
Add a call to kill_child in case of early return after fork.
(linux_check_ptrace_features): Use kill_child instead of local code.
(linux_test_for_tracefork): Likewise.
This removes an IMO not very useful comment in gdb/Makefile.in about
"alloca". It also removes INFOFILES, which I think probably has not
been useful since whenever the manual was moved into a subdirectory.
gdb/ChangeLog
2018-12-14 Tom Tromey <tom@tromey.com>
* Makefile.in: Remove "alloca" comment.
(INFOFILES): Remove.
(local-maintainer-clean): Don't use INFOFILES.
Linkers group input note sections with the same name into one output
note section with the same name. One output note section is placed in
one PT_NOTE segment. New linkers merge all input .note.gnu.property
sections into one output .note.gnu.property section with a single
NT_GNU_PROPERTY_TYPE_0 note in a single PT_NOTE segment. Since older
linkers treat input .note.gnu.property section as a generic note section
and just concatenate all input .note.gnu.property sections into one
output .note.gnu.property section without merging them, we may
see one or more NT_GNU_PROPERTY_TYPE_0 notes in PT_NOTE segment, which
are invalid.
GNU_PROPERTY_X86_UINT32_VALID was defined to address this issue such
that linker sets the bit for non-relocatable outputs. But it isn't
sufficient:
1. It doesn't cover generic properties.
2. When -mx86-used-note=yes is passed to x86 assembler, the
GNU_PROPERTY_X86_UINT32_VALID bit is set in GNU_PROPERTY_X86_ISA_1_USED
property in object file and older linkers generate invalid
NT_GNU_PROPERTY_TYPE_0 notes with the GNU_PROPERTY_X86_UINT32_VALID bit
set.
I am proposing the following changes:
1. Add PT_GNU_PROPERTY segment type:
# define PT_GNU_PROPERTY (PT_LOOS + 0x474e553)
which covers .note.gnu.property section.
2. Remove GNU_PROPERTY_X86_UINT32_VALID.
bfd/
PR ld/23900
* elf.c (get_program_header_size): Add a PT_GNU_PROPERTY
segment for NOTE_GNU_PROPERTY_SECTION_NAME.
(_bfd_elf_map_sections_to_segments): Create a PT_GNU_PROPERTY
segment for NOTE_GNU_PROPERTY_SECTION_NAME.
* elfxx-x86.c (_bfd_elf_link_setup_gnu_properties): Don't set
GNU_PROPERTY_X86_UINT32_VALID.
binutils/
PR ld/23900
* readelf.c (get_segment_type): Support PT_GNU_PROPERTY.
(decode_x86_isa): Don't check GNU_PROPERTY_X86_UINT32_VALID.
(decode_x86_feature_1): Likewise.
(decode_x86_feature_2): Likewise.
(print_gnu_property_note): Remove GNU_PROPERTY_X86_UINT32_VALID
check.
* testsuite/binutils-all/i386/empty.d: Updated.
* testsuite/binutils-all/x86-64/empty-x32.d: Likewise.
* testsuite/binutils-all/x86-64/empty.d: Likewise.
* testsuite/binutils-all/i386/pr21231b.s: Change
GNU_PROPERTY_X86_ISA_1_USED bits to 0x7fffffff.
* testsuite/binutils-all/x86-64/pr21231b.s: Likewise.
gas/
PR ld/23900
* config/tc-i386.c (x86_cleanup): Don't set
GNU_PROPERTY_X86_UINT32_VALID.
* testsuite/gas/i386/property-1.s: Change
GNU_PROPERTY_X86_ISA_1_USED bits to 0.
include/
PR ld/23900
* elf/common.h (PT_GNU_PROPERTY): New.
(GNU_PROPERTY_X86_UINT32_VALID): Removed.
ld/
PR ld/23900
* testsuite/ld-elf/elf.exp: Run PR ld/23900 test.
* testsuite/ld-elf/pr23900-1-32.rd: New file.
* testsuite/ld-elf/pr23900-1-64.rd: Likewise.
* testsuite/ld-elf/pr23900-1.d: Likewise.
* testsuite/ld-elf/pr23900-1.s: Likewise.
* testsuite/ld-elf/pr23900-2.s: Likewise.
* testsuite/ld-elf/pr23900-2a.d: Likewise.
* testsuite/ld-elf/pr23900-2b.d: Likewise.
* testsuite/ld-i386/ibt-plt-1.d: Adjusted.
* testsuite/ld-i386/ibt-plt-2c.d: Likewise.
* testsuite/ld-i386/ibt-plt-2d.d: Likewise.
* testsuite/ld-i386/ibt-plt-3d.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-1-x32.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-1.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-2c-x32.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-2c.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-2d-x32.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-2c.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-3c-x32.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-3c.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-3d-x32.d: Likewise.
* testsuite/ld-x86-64/ibt-plt-3d.d: Likewise.
* testsuite/ld-i386/pr23372c.d: Expect <None>
for GNU_PROPERTY_X86_ISA_1_USED.
* testsuite/ld-x86-64/pr23372c-x32.d: Likewise.
* testsuite/ld-x86-64/pr23372c.d: Likewise.
* testsuite/ld-x86-64/pr23372d-x32.d: Likewise.
* testsuite/ld-x86-64/pr23372d.d: Likewise.
* testsuite/ld-x86-64/property-x86-5a.s: Change
GNU_PROPERTY_X86_ISA_1_USED bits to 0.
* testsuite/ld-x86-64/property-x86-5b.s: Likewise.
Add a script to generate the FreeBSD XML system call table from the
sys/sys/syscall.h file in the kernel source tree. For ABI
compatiblity system calls used by older binaries (such as
freebsd11_kevent()), the original system call name is used as an
alias.
Run this script against the current syscall.h file in FreeBSD's head
branch which is expected to be the file used in 12.0 (head is
currently in code freeze as part of the 12.0 release process).
gdb/ChangeLog:
* syscalls/update-freebsd.sh: New file.
* syscalls/freebsd.xml: Regenerate.
When setting a syscall catchpoint by name, catch syscalls whose name
or alias matches the requested string.
When the ABI of a system call is changed in the FreeBSD kernel, this
is implemented by leaving a compatibility system call using the old
ABI at the existing "slot" and allocating a new system call for the
version using the new ABI. For example, new fields were added to the
'struct kevent' used by the kevent() system call in FreeBSD 12. The
previous kevent() system call in FreeBSD 12 kernels is now called
freebsd11_kevent() and is still used by older binaries compiled
against the older ABI. The freebsd11_kevent() system call can be
tagged with an "alias" attribute of "kevent" permitting 'catch syscall
kevent' to catch both system calls and providing the expected user
behavior for both old and new binaries. It also provides the expected
behavior if GDB is compiled on an older host (such as a FreeBSD 11
host).
gdb/ChangeLog:
* NEWS: Add entry documenting system call aliases.
* break-catch-syscall.c (catch_syscall_split_args): Pass 'result'
to get_syscalls_by_name.
* gdbarch.sh (UNKNOWN_SYSCALL): Remove.
* gdbarch.h: Regenerate.
* syscalls/gdb-syscalls.dtd (syscall): Add alias attribute.
* xml-syscall.c [!HAVE_LIBEXPAT] (get_syscalls_by_name): Rename
from get_syscall_by_name. Now accepts a pointer to a vector of
integers and returns a bool.
[HAVE_LIBEXPAT] (struct syscall_desc): Add alias member.
(syscall_create_syscall_desc): Add alias parameter and pass it to
syscall_desc constructor.
(syscall_start_syscall): Handle alias attribute.
(syscall_attr): Add alias attribute.
(xml_get_syscalls_by_name): Rename from xml_get_syscall_number.
Now accepts a pointer to a vector of integers and returns a
bool. Add syscalls whose alias or name matches the requested
name.
(get_syscalls_by_name): Rename from get_syscall_by_name. Now
accepts a pointer to a vector of integers and returns a bool.
* xml-syscall.h (get_syscalls_by_name): Likewise.
gdb/doc/ChangeLog:
* gdb.texinfo (Set Catchpoints): Add an anchor for 'catch syscall'.
(Native): Add a FreeBSD subsection.
(FreeBSD): Document use of system call aliases for compatibility
system calls.
This removes the need for the caller to explicitly manage the memory
for the returned system call list. The sole caller only needed the
system call numbers rather than the full syscall structures.
get_syscalls_by_group now uses a boolean return value to indicate if
the requested group exists.
gdb/ChangeLog:
* break-catch-syscall.c (catch_syscall_split_args): Pass 'result'
to get_syscalls_by_group.
* xml-syscall.c [!HAVE_LIBEXPAT] (get_syscalls_by_group): Return
false.
[HAVE_LIBEXPAT] (xml_list_syscalls_by_group): Append syscall
numbers to an existing vector of integers and return a bool.
(get_syscalls_by_group): Accept pointer to vector of integers
and change return type to bool.
* xml-syscall.h (get_syscalls_by_group): Likewise.
* riscv-tdep.c (riscv_print_one_register_info): For MSTATUS, add
comment for SD field, and correct xlen calculation. For MISA, add
comment for MXL field, add call to register_size, and correct base
calculation.
2018-12-12 Philippe Waroquiers <philippe.waroquiers@skynet.be>
* nat/linux-osdata.c (linux_xfer_osdata_info_os_types): Forward
declare on one line to fix ARI warning.
When running the test gdb.base/annota1.exp with:
make check-gdb RUNTESTFLAGS="--target_board=native-extended-gdbserver gdb.base/annota1.exp"
I would see a failure due to some unexpected lines in GDB's output.
The extra lines (when compared with a native run) were about file
transfer from the remote back to GDB.
This commit extends the regexp for this test to allow for these extra
lines, and also splits the rather long regexp up into a list of parts.
With this change in place I see no failures for gdb.base/annota1.exp
when using the native-extended-gdbserver target board, nor with a
native run on X86-64/Linux.
gdb/testsuite/ChangeLog:
* gdb.base/annota1.exp: Update a test regexp.
I ran into a situation where attempting to make an inferior function
call would trigger an assertion, like this:
(gdb) call some_inferior_function ()
../../src/gdb/regcache.c:310: internal-error: void regcache::restore(readonly_detached_regcache*): Assertion `src != NULL' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
Quit this debugging session? (y or n)
The problem that triggers the assertion is that in the function
save_infcall_suspend_state, we basically did this:
1. Create empty infcall_suspend_state object.
2. Fill fields of infcall_suspend_state object.
The problem is causes is that if filling any of the fields triggered
an exception then the infcall_suspend_state object would be deleted
while in a partially filled in state.
In the specific case I encountered, I had a remote RISC-V target that
claimed in its target description to support floating point registers.
However, this was not true, and when GDB tried to read a floating
point register the remote sent back an error. This error would cause
an exception to be thrown while creating the
readonly_detached_regcache, which in turn caused GDB to try and delete
an infcall_suspend_state which didn't have any register state, and
this triggered the assertion.
To prevent this problem we have two possibilities, either, rewrite the
restore code the handle partially initialised infcall_suspend_state
objects, or, prevent partially initialised infcall_suspend_state
objects from existing. The second of these seems like a better
solution.
So, in this patch, I move the filling in of the different
infcall_suspend_state fields within a new constructor for
infcall_suspend_state. Now, if generating one of those fields fails
the destructor for infcall_suspend_state will not be executed and GDB
will not try to restore the partially saved state.
With this patch in place GDB now behaves like this:
(gdb) call some_inferior_function ()
Could not fetch register "ft0"; remote failure reply 'E99'
(gdb)
The inferior function call is aborted due to the error.
This has been tested against x86-64/Linux native, native-gdbserver,
and native-extended-gdbserver with no regressions. I've manually
tested this against my baddly behaving target and confirmed the
inferior function call is aborted as described above.
gdb/ChangeLog:
* infrun.c (infcall_suspend_state::infcall_suspend_state): New.
(infcall_suspend_state::registers): New.
(infcall_suspend_state::restore): New.
(infcall_suspend_state::thread_suspend): Rename to...
(infcall_suspend_state::m_thread_suspend): ...this.
(infcall_suspend_state::registers): Rename to...
(infcall_suspend_state::m_registers): ...this.
(infcall_suspend_state::siginfo_gdbarch): Rename to...
(infcall_suspend_state::m_siginfo_gdbarch): ...this.
(infcall_suspend_state::siginfo_data): Rename to...
(infcall_suspend_state::m_siginfo_data): ...this.
(save_infcall_suspend_state): Rewrite to use infcall_suspend_state
constructor.
(restore_infcall_suspend_state): Rewrite to use
infcall_suspend_state::restore method.
(get_infcall_suspend_state_regcache): Use
infcall_suspend_state::registers method.
